Grandfathers Matter(ed): Occupational Mobility Across Three Generations in the U.S. and Britain, 1850-1911

D EPT

Jason Long OF B USINESS & E CONOMICS W HEATON C OLLEGE

AND

Joseph Ferrie D EPT OF E CONOMICS N ORTHWESTERN U NIVERSITY AND NBER

Nearly all intergenerational mobility studies focus on fathers and sons. The possibility that the process is more than simply two-generational [AR(1)] has been difficult to assess because of the lack of the necessary multi-generational data. We remedy this shortcoming with new data that links grandfathers, fathers, and sons in Britain and the U.S. between 1850 and 1910. We find that grandfathers mattered: even controlling for father’s occupation, grandfather’s occupation significantly influenced the occupation of the grandson. For both Britain and the U.S. in this time period, therefore, assessments based on two-generation estimates significantly overstate the true amount of social mobility.

The transmission of economic and social outcomes such as earnings, occupation, and education across generations has long interested a wide range of social scientists. Intergenerational social mobility has been a central topic of empirical sociology for many years. More recently, measuring the intergenerational elasticity of earnings has received a great deal of attention in economics, particularly following the publication more than twenty years ago of influential papers by Gary Solon (1992) and David Zimmerman (1992). One commonality across this large and diverse literature is the dominance of studies that analyse only two generations. Solon (1999) and Black and Devereux (2011) provide broad surveys of the empirical economics studies, with an emphasis on the intergenerational elasticity of earnings. Björklund and Salvanes (2011) survey the equally extensive literature on the impact of family background on educational outcomes, including studies of direct parental effects and of sibling correlations. The overwhelming majority of the studies surveyed from these literatures focus exclusively on the transmission of outcomes from parent to child and do not consider the potential influence of grandparent and further-

–1–

removed generations.1 The theoretical work that informs the empirical studies, particularly the Becker-Tomes model on which so many of the studies rely, is a two-generation model in which the child’s outcome is a function of investments made by and an endowment received from the parents (Becker and Tomes, 1979).2 This focus on the impact of parents on children is not so much one of choice as of necessity. Typically, the available information covers only two generations, whether the data source follows individuals over time or has retrospective questions on an individual’s early-life household. In fact, the overarching goal of economic and social mobility studies is to determine how family background – in the sense of a broad-ranging familial endowment of genes, investments in human capital, social networks and the like – influences individuals’ educational and labour market prospects. In theory, there is no reason why this should be a simple two-generation AR(1) process. 3 It is entirely plausible that the impact of family background characteristics goes back farther than just the parental generation. Whether it does is an empirical question, and if it does, then the large literature that measures and compares intergenerational mobility across many countries and time periods systematically overestimates true mobility rates by assuming that only the previous generation matters for the prospects of the current one. Sociologists have explored the possible effect of multiple generations more than have economists. Warren and Hauser (1997) provide a useful survey of sociological work on mobility over multiple generations, some of which has found evidence for multigenerational effects and some of which has not. They analyse the Wisconsin Longitudinal Study, which includes information over a long time period on a sample of 1957 Wisconsin high school graduates. Approximately 4,000 individuals from the sample report occupational information for their parents and at least one of their children over the course of their study. The authors find that the occupation of the grandfather does not influence the occupation of the grandson if the occupation of the father is controlled for. An obvious limitation of this study is its reliance on data from only one state; whether this finding applies to the U.S. as a whole cannot be determined. For the most part, economists are just starting to examine the multigenerational transmission of occupation, earnings, or education. In an early study of the intergenerational elasticity of earnings between fathers and sons, Peters (1992) includes grandfather’s education as a control variable in her estimation equation, and finds no significant effect.

                                                             Though some studies consider mothers and daughters, the analysis of men is far more typical. In what follows, we will refer to grandfathers, fathers, and sons in the interest of brevity and clarity, though all of the analytical framework could just as well refer to females in any of the generations considered. 2 This basic structure is also followed by Solon (2004). 3 That is, a temporal process in which only the previous period directly influences the present period. 1

–2–

However, her data does not reveal grandfathers’ earnings, nor is the analysis of multigenerational effects an explicit aim of her study. More recently, several studies have explored multigenerational mobility in a variety of mostly European contexts. A team of Swedish researchers has conducted a detailed analysis of the persistence of human capital over four generations of individuals originating with the 1938 Malmö Study (Lindahl et al., 2014, 2015). The original study surveyed 1,542 third graders in the Malmö metropolitan area in 1938. Since that time, subsequent generations have been added to the dataset so that at present the authors are able to observe earnings for three generations and education for a fourth for a total of 901 complete families. They find that simple two-generational estimates of the elasticity of earnings and education between parent and child significantly underpredict the true persistence across three generations. This finding runs counter to the finding of Warren and Hauser for Wisconsin and indicates that multigenerational effects matter for this particular sample. Like the Warren and Hauser study, and like many of the studies they survey from the sociology literature, the dataset used contains relatively few observations across all three generations and is not clearly nationally representative. Braun and Stuhler (2016) analyse data from several retrospective surveys in Germany. Like Lindahl et al., they find that long-run social mobility is lower than the rate suggested by two-generations of data alone. Dribe and Helgertz (2016) find a similar result for nineteenth- century Sweden with respect to the multigenerational transmission of occupational status, but find no such relationship when they examine earnings. In addition to these specific studies, Solon (2015) provides a useful review of additional work on multigenerational mobility, particularly some significantly older studies, and a discussion of different theoretical interpretations of the empirical findings. Several recent studies have considered mobility and the persistence of inequality over very long runs of family and dynastic history using methodology quite different from the standard intergenerational regression framework. Clark and Cummins (2015) link seven generations of families with rare surnames in England from 1800 to the present. Rather than comparing income and education between individual parents and their children, they compare average wealth of surname-sharing families across generations. With this methodology, they are able to include up to five generations in the wealth estimation equation for their most recent generation. Their results indicate that, while the impact of each successive generation diminishes, all five generations do exert a significant influence, leading the authors to conclude that existing studies significantly overestimate the true rate of mobility. Clark (2014) and Cummins and Clark (2015) similarly use innovative surname data over the very long run to demonstrate that the true rate of social mobility is higher than has previously been thought. Studies based on archival data on the descendants of Qing Dynasty emperors in China also indicate a far-reaching effect of ancestry on

–3–

the occupational status of subsequent generations (Campbell and Lee, 2011; Mare and Song, 2012). Whether multiple generations do in fact influence the occupational attainment of individuals is an empirical question, but there are good reasons to expect, a priori, that they could. Mare (2011) describes several mechanisms by which multigenerational influence could operate. The accumulation within a family of sufficient wealth in the form of financial or physical capital would be one channel by which generations prior to the parent could influence the outcomes of children. The availability and quality of a wider kin network that could assist with child rearing, job acquisition, et cetera is another. One might extend this concept to include the relevant social capital available to individuals, the accumulation of which could be influenced by generations prior to the parents. Finally, various biological mechanisms could determine the inheritance of salient traits across more than two generations.

Empirical Framework The implicit assumption in much of the work on intergenerational mobility is that the transition from fathers to sons is independent of the history of previous father-son transitions in the same family line. In terms of income or wealth, ln

ln

(1)

,

where Yit is the outcome for an individual in family line i in generation t, Yit-1 is the corresponding outcome for another individual in family line i in generation t-1, and εit is an error term with the usual properties. β1 represents father-son intergenerational income/wealth elasticity (IGE). Under this assumption, Y is the outcome of a simple AR(1) process, so if the term Yit-1 is replaced with Yit-2: ln

ln

(2)

,

the grandfather-grandson IGE, β2, would simply be the square of β1. If the process is not in fact AR(1), if prior generations exert a greater influence on current outcomes than simply through their influence on fathers, then we would observe β2 > (β1)2. Alternately, if we consider instead the intergenerational transmission of occupational class or status, the occupational distribution of generation t is determined according to Occupations

Occupations

–4–

′

(3)

where the first term is a vector of 1,...,m occupational counts in generation t, the second is the corresponding vector in generation t-1, and M1 is an m × m matrix of transition probabilities. Again, under the assumption that only one previous generation matters for occupational class attainment, an individual’s occupational category in generation t is the outcome of a Markov process, so outcomes for generation t are related to those in generation t-2 by Occupations

Occupations

′

(4)

and M2 is the square of the Markov matrix M1. If on the other hand prior generations matter, then [Occupationst-2]M1M1 will exhibit more mobility than is actually observed between grandfathers and grandsons. It follows then that the simplest way to assess whether intergenerational mobility is described by a process that ignores any history prior to generation t-1 is to see whether in fact β2 = (β1)2 for income or wealth and whether [Occupationst-2]M2 = [Occupationst2]M1M1.

Both methods, of course, require observation of adult labour market outcomes

for three generations. In this paper, we use new data on grandfathers, fathers, and sons linked across nineteenth and early twentieth century censuses in the U.S. and Britain to assess the degree of mobility across three generations in both countries. The data are constructed from nationally-representative sources in a base year (the 1881 British Census and the 1880 U.S. Census) which are then linked to (1) the appearance of the fathers from those sources (when they were children residing with own fathers) in the census three decades prior; and (2) the appearance of the sons from those sources (when they were adults) in the census three decades subsequent. This permits an analysis of mobility across three generations in each country and a characterization of the differences in those patterns across two countries for which we have found substantial differences in two generation mobility in previous work (Long and Ferrie, 2013).

Linking Generations Across Historical Censuses Previous work on intergenerational mobility has, for the most part, relied on simple comparisons of fathers and sons out of necessity – suitable, nationally representative datasets providing information on three or more generations has not been readily available, with the recent exceptions mentioned above. The recent completion of indexes to the British (through 1911) & U.S. (through 1940) population censuses allows us to generate such samples by linking three generations across censuses for both countries.

–5–

British Data We used three sources to construct the data for Britain4: (1) a computerised two percent sample of the 1851 census, (2) a computerised version of the complete count 1881 census, and (3) the complete count 1911 census, accessible through Ancestry.com, a Web-based genealogical research service.5 The first stage of the data creation process was the nominal linkage of 28,474 men from the 1851 census sample to the 1881 complete-count census.6 Individuals from the 1851 census were identified in the 1881 census based on first and last name and county, parish and year of birth – information that should, barring error, remain constant across censuses. Some leeway in the matching algorithm was allowed for small discrepancies in reporting personal information across censuses thirty years apart. First and last names were required to match according to the SOUNDEX algorithm, which is designed to match English language words phonetically, and middle initial had to be consistent. Birth year was allowed to deviate by up to five years from the expected value based on reported age in 1851. When multiple records from 1881 satisfied the linkage criteria for a single target individual from 1851, that individual was dropped. This matching algorithm was applied to an initial pool of 168,130 men from the 1851 census, yielding a success rate of 17 percent. Using age-specific mortality rates (Mitchell, 1962, p. 38-39) and emigration rates (Baines, 1985, p. 152-153), we estimate that approximately 85,000 of the males in the initial sample would be expected to die by 1881, and that 13,500 would be expected to migrate out of England and Wales. The ideal expected match success rate, then, is 41 percent. Part of the discrepancy comes from duplicate matching: we eliminated 9,622 individuals from the 1851 sample because they matched to more than one person in the 1881 census. The gap between the expected 41 percent success rate and the 23 percent of individuals who actually were matched (uniquely or not) must be attributed to enumeration error – the individuals misreported their age by more than five years, or the birthplace or name information given in the two censuses could not be reconciled, or they simply were not enumerated in 1881. This degree of match success is in line with other studies that rely on linked census data according to a recent review of linkage techniques (Bailey et al., 2017).7                                                              Throughout the paper, we use the term “Britain” as a matter of convenience. In fact, the data include only information on individuals residing in England and Wales. When the project was begun, Scotland was not included in the 1881 data source. The Scottish data are now available, and will eventually be incorporated into the analysis. 5 The 1851 and 1881 data sets are available from the UK Data Archive, as studies number 1316 and 4177, respectively. 6 This 1851-1881 linked census dataset was first described in Long (2005), from which the data description here draws. 7 To test the robustness of the matching procedure, the intergenerational mobility analysis was repeated using only the individuals who were matched exactly between 1851 and 1881 – in other words, those individuals who reported precisely the same name in both censuses and whose reported age in 1881 was exactly 30 4

–6–

Because duplicate matches are dropped, individuals with the most common names are less likely to be included in the sample. We might expect such individuals more frequently to come from the lower socioeconomic strata, and this might impart bias to the matched sample. To test for this possibility, we use all of the males in the 1851 census sample to estimate a probit regression in which the dependent variable is whether an individual’s name occurs eight times or more in the 1851 sample and the explanatory variables are age, county of residence, and dummy variables indicating occupational status as labourer, farmer, or servant.8 Age and the occupational dummies are statistically significant, but none is practically significant; that is, the marginal effects are all small. The largest coefficient is for the servant dummy, which increases the probability of having a common name by only 1.6 percentage points, relative to a baseline probability of 25 percent. So the “common name” characteristic should not impart a troublesome degree of bias to the matched sample.9 Because the census records households together and because the 1851 sample preserves this household structure, it is simple to connect the young linked males who were sons living with their family in 1851 with their fathers. The matched data contain 12,647 such father/son pairs. The fathers are generation one (G1) in this study, and the sons are generation two (G2). This provides the basic structure necessary to observe intergenerational mobility from 1851 to 1881, and it does so in such a way that both father and son are observed as mature adults, with approximately equal ages, at approximately the same point in the life cycle. For the 12,647 father/son pairs in which the son is aged 0 – 19 years and living with his father in 1851, the average age of the father in 1851 is 41.5 years, and the average age of the son thirty years later in 1881 is 38.0 years. An average age difference of only 3.5 years should have a negligible impact on observed mobility considering the advanced age of both father and son at the time of observation.10 The final stage of the linkage process adds a third generation to the data. 20,269 sons were extracted from the households of the linked males in 1881, and 6,672 were uniquely identified in the 1911 complete-count census using Ancestry.com’s Web-based genealogical                                                              years greater than their reported age in 1851. The mobility pattern for this group is essentially the same as the results reported for the data as a whole. The Altham G2 test statistic, described in a following section, shows no statistically significant difference in mobility between the exact-match subset and the whole data set. 8 I define “common” here so that 25 percent of the population has a common name; the eight-occurrence cutoff follows from this definition. Results are unchanged if the definition is changed to 10 percent of the population. Excluding the county variable also does not change the results. 9 This finding is consistent with other studies using matched data that also fail to find evidence of a “common name” problem (Steckel, 1988; Ferrie, 1999, p. 20-31). For more details on the linkage procedure and data construction process, see Long and Ferrie (2013) and Long (2005). 10 The age structure of the linked census data makes it particularly well suited to measuring intergenerational mobility in nineteenth century England relative to the sources that have typically been used in the past. For more, see Long (2013).

–7–

research service. These records were linked one-by-one using Ancestry’s own search function. Though the search function is proprietary and its specifics are not published, it is a high precision linkage method that blends standard computerized linkage algorithms with verified individual genealogical linkages created by family history researchers. The match success rate of 33 percent is higher than the success rate of the 1851-1881 linkage and is fairly high relative to other studies, a fact most likely due to the accuracy of Ancestry’s internal search function. This is generation three (G3), used to measure intergenerational mobility from 1881 to 1911. In all, we observe 6,174 three-generation lineages. Of these, we restrict attention to the 3,975 in which each of G1, G2, and G3 is between the ages of 30-59 – therefore clearly holding an adult occupation – and for whom we observe classifiable occupation information. Table 1 uses the households of three successive generations of males from 1851, 1881, and 1911 to illustrate the nature of the linked census data. The table is also illustrative of the nature of the U.S. data. We use both observed occupation and earnings imputed by occupational class as labour market outcomes to assess mobility. The censuses only directly reveal occupation. Although earnings is the measure most commonly used in the economics literature on mobility, there are advantages to using occupation (as is the norm in the sociology literature). One of the principal empirical difficulties in the study of earnings mobility is obtaining a true measure of permanent income in the face of frequent transitory income shocks. As shocks often occur without job changes, occupation should be less affected by such disturbances. Further, compared to a simple earnings measure, occupation and social class capture more dimensions of an individual’s experience that may be related to interpretations of social mobility, such as prestige in the community, autonomy in the workplace, manual versus non-manual labour, place of work, and so on. We code the British occupations according to two different classification schemes. First, we group all occupations into four broad comparisons – white collar, farmer, skilled & semi-skilled, and unskilled – that are defined consistently across censuses. We have used this system elsewhere in order to make comparisons across Britain and the U.S. (Long and Ferrie, 2013). It offers the advantage of being equally appropriate for both economies, even considering their significant differences in labour market structure. Its primary disadvantages are that the groups are quite coarse, and that, for the most part, the occupation groups are not clearly rankable according to typical earnings and/or socioeconomic status.11 With this scheme, we only unambiguously rank unskilled labour as the

                                                             11 The main difficulty in ranking occupational groups has to do with farmers in the U.S., who range from very high-earning large landowners to poor farmers with small and/or marginal holdings who earned less than skilled labourers. Farmers are much easier to confidently rank (highly) in Britain, where there are far fewer of them and they earn typically much more relative to other occupations.

–8–

TABLE 1 EXAMPLE OF MULTI-GENERATIONAL LINKED CENSUS DATA, BRITAIN, 1851–1911 1851 Census: Phillips household; Eastergate parish, Sussex, England County Parish

First Name

Last Name

Rel

Mar Occupation

Age

Sussex Eastergate William Phillips Head Sussex Eastergate Martha Phillips Wife

M Ag Labourer M -

Sussex Eastergate William Phillips Son Sussex Eastergate Mary Phillips Daur Sussex Eastergate Richard Rewell Lodger

U Ag Labourer U Scholar W Ag Labourer

42 37 17 13 85

Birth Birth County Parish Sussex Chichester Sussex Walberton Sussex Walberton Sussex W Hampnet Sussex Walberton

1881 Census: Phillips household; Arundel parish, Sussex, England County Parish

First Name

Sussex Arundel Sussex Arundel Sussex Arundel

William Phillips Head Jane Phillips Wife George Phillips Son

M Blacksmith M Bsmith wife U Bricklayers lab

Sussex Arundel Sussex Arundel

David Phillips Son Thomas Phillips Son

U Scholar U -

Last Name

Rel

Mar Occupation

Age

Birth Birth County Parish

47 Sussex Walberton 45 Sussex Arundel 15 Sussex Arundel 8 Sussex Arundel 10m Sussex Arundel

1911 Census: Phillips household; Arundel parish, Sussex, England County Parish

First Name

Last Name

Rel

Sussex Sussex Sussex Sussex

David Emily Annie Emily

Phillips Phillips Phillips Phillips

Head Wife Daur Daur

Arundel Arundel Arundel Arundel

Mar Occupation M Gardener M U U -

Age 38 31 10 8

Birth Birth County Parish Sussex Hull Sussex Sussex

Arundel Yorkshire Arundel Arundel

Note : Three linked individuals used to measure intergenerational mobility shown in italics. For the sake of clarity, some members of each household are not shown.

lowest category; upward and downward mobility, therefore, are defined as mobility out of or into unskilled labour, respectively. For analysis of the British data where comparison with the U.S. is not necessary, we use W. A. Armstrong’s ordered classification system, which is based on the Registrar General’s 1921 and 1951 classification schemes (Armstrong, 1972). Every individual is assigned to one of five ranked social classes according to his occupation as recorded in the census enumerator’s book: I – Professional, II – Intermediate, III – Skilled, IV – Semiskilled, and V – Unskilled.12 This is a classification system based solely on occupation, and while

                                                             Some typical occupations are Class I – solicitor, accountant; Class II – farmer, carpenter (employer); Class III – carpenter (not employer), butcher (not employer), skilled in manufacturing; Class IV – agricultural labourer, wool comber; Class V – general labourer, porter. 12

–9–

there surely are additional components of social class, occupation is nearly always considered to be of central importance in determining an individual’s class. Armstrong’s aim for the classification system is to “ensure that each category is homogeneous in relation to the basic criterion of the general standing within the community of the occupations concerned.” Under Armstrong’s system, each occupation is coded according to the Registrar General’s classification scheme, with several modifications made to minimise anachronism. The most important modification is that, regardless of job title, all employers of 25 or more are included in Class I, and all individuals with Class III or IV occupations employing at least one person other than a family member are included in Class II. Empirically, this scheme correlates well with other indicators of social class. In this sense, what is referred to here as “social class” is essentially synonymous with what is often referred to in the social sciences as socioeconomic status. According to occupational/industrial pay estimates compiled by Jeffrey Williamson (1980, 1982), the average wage premium in 1851 for each class relative to the next lowest class was 7 percent for Class IV, 33 percent for Class III, 81 percent for Class II, and 45 percent for Class I. Furthermore, Armstrong (1972, p. 212) demonstrates that job class, defined according to this system, is positively correlated with the employment of servants and negatively correlated with the incidence of shared accommodation. While earnings is not revealed in the census, it can be imputed to individuals based on their reported occupations. While no source is ideally suited to the task, Williamson’s pay estimates serve reasonably well. There are well-known problems with these estimates, primarily having to do with a handful of professional occupations, especially solicitors and barristers, surgeons and doctors, and engineers. The wage information for these particular occupations is derived from a small number of sources and demonstrates extreme variation across time periods (Jackson, 1987; Feinstein, 1988). For the limited purposes of this study, these problems can be dealt with simply enough by omitting these occupations in the construction of the average wage for each class and year. This method preserves some of the important advantages of the Williamson wage information relative to other sources, particularly its consistent construction across most of relevant timespan of the data.13 The occupations used for each class along with the average wage for each relevant year are shown in Table 2. Wages are imputed by specific occupation for 72

                                                             Another advantage is the inclusion of some wage information for white-collar workers. While the previously mentioned professional occupations are problematic, there are a handful of occupations included with which to derive an average wage for classes I and II. Other viable sources for occupational wages from one specific time period, such as Leone Levi’s 1885 report, lack any wage information for the professional occupations (Levi, 1885). 13

– 10 –

TABLE 2 CLASS AND INCOME FOR COMMON BRITISH OCCUPATIONS Class I: Professional

Occupations for 1851-1901 High-wage gov employee, Clergy

II: Intermediate Clerk, Teacher

1851

1881

1901

£250.98 £295.33 £198.82 158.46

203.73

217.18

III: Skilled

Skilled worker in engineering, building, shipbuilding and textiles

66.92

85.69

95.8

IV: Semiskilled

Farm labourer, Miner, Railway worker, Low-wage gov employee

48.02

58.65

69.26

V: Unskilled

General non-agricultural labourer

44.83

55.88

68.9

Sources : Williamson (1980, 1982)

percent of the individuals in the data; the rest are assigned the average wage for their Armstrong occupational class.14 We refer to this variable as “occupational earnings”. A critical issue with linked census data is the extent to which it preserves the representative nature of the underlying data sources. To assess this question, we compare each generation of the linked data to the population in that year. As Table 3 shows, the linkage procedure we employ produces a dataset that is well representative of the male population of England and Wales, conditional on age and fatherhood status.15 The population reference data comes from two-percent random samples drawn from the 1851, 1881, and 1911 censuses of England and Wales.16 The linked data are compared to the men from these samples who were aged 30-59 and had at least one child co-resident in the household at the time of enumeration. Across the dimensions observable from the censuses, the linked data fairly closely resembles the population reference data. In particular, the occupational class and especially the occupational earnings of the men in the two samples are quite similar, indicating that, for the most part, the linkage procedure did not introduce a significant economic bias into the data. Of course, this multigenerational data will not be perfectly representative of even the similarly aged fathers in 1851. To be in our dataset, an

                                                             14 Williamson’s data covers the years 1851, 1881, and 1901, but does not extend to 1911; therefore, the 1901 wage information is used to impute wages for Generation 3 in 1911. 15 For a more detailed representativeness analysis of just the 1851-1881 linkage, see Long (2005, p. 5-7). 16 This census data is available via the North Atlantic Population Project (Minnesota Population Center, 2015). Specific datasets used are National Sample from the 1851 Census of Great Britain (Schürer and Woollard, 2008), National Sample from the 1881 Census of Great Britain (Schürer and Woollard, 2003), and National Sample from the 1911 Census of Great Britain (Schürer and Woollard, 2014).

– 11 –

TABLE 3 REPRESENTATIVENESS OF LINKED SAMPLE, BRITAIN 1851 1881 1911 Linked Pop Linked Pop Linked Pop

Age (years) Number of children Occ Earnings (₤) Occ Class (%) I. Prof II. Intermediate III. Skilled IV. Semi-S V. Unskilled Region (%) East Lan-Chs London London Environs Midlands North South Wales York N

43.6 4.5 61.0

42.4 3.3 64.9

40.7 4.4 86.8

42.3 3.5 85.0

39.1

42.5

104.9

103.6

2.2 16.5 41.5 30.5 9.4

1.8 11.8 50.5 24.8 11.2

2.9 15.2 49.0 21.4 11.5

1.9 10.7 57.2 14.8 15.4

1.7 13.8 57.5 12.2 15.0

1.7 11.9 59.5 9.9 17.0

11.1 15.7 4.0 14.8 16.2 7.8 20.9 2.2 7.5

6.6 14.5 8.8 10.8 19.3 5.8 19.6 5.5 9.1

9.9 16.7 7.3 14.0 15.1 8.4 19.3 2.5 6.9

3.8 16.2 10.7 13.6 18.3 6.7 13.8 5.2 11.7

5,129

31,730

6,094

48,962

6,010

77,380

Note: Representative data are men age 30-59 with at least one co-resident child from 2% random samples drawn from the 1851, 1881, and 1911 censuses of England and Wales. Number of children reflects only those children co-resident at the time of enumeration. See text for definition of Occupational Earnings and Occupational Class. Number of children and region of residence were not transcribed from the 1911 census linkage through Ancestry.com.

individual in 1851 must not only have a male child, but also that child must survive to adulthood and have his own child who survives to adulthood. A straightforward implication of this feature of the data is that men in our linked sample will have more children on average, simply because it increases the likelihood that at least one of their children will satisfy this three-generation linkage criteria. It is also natural that the linked men would be younger on average in 1881 and 1911, as these were the two generations that had to survive and remain in Britain for thirty years from their first census in order to be linked.

– 12 –

U.S.Data The U.S. data reports the occupation of a male household head in 1850 (Generation 1), the occupation of his son in 1880 (Generation 2), and the occupation of the son of the Generation 2 male in 1910 (Generation 3). The linkages across three U.S. population censuses proceeded in three steps: 1. All males present in the Full-Count File of the 1850 U.S. Population Census who met the following criteria were extracted: (1) they were age 3-21 in 1850; and (2) they were co-resident with both parents in 1850. A total of 3,057,484 individuals satisfied both criteria. The age requirement was imposed so they would be independent household heads thirty years later in 1880. The co-residence requirement was imposed to determine the birthplaces of both parents (which was necessary in the next step) and, in the case of fathers, to determine the father’s occupation. Requirement (2) necessarily creates a bias toward intact households which, for a variety of reasons, will be unrepresentative of the total population of males age 3-21 in 1850.17 (As the same requirement will be imposed in subsequent stages of the linkage process, however, this bias will be consistent throughout the linked sample.) 2. The individuals extracted from the 1850 census in Step (1) were then sought in the Full-Count File of the 1880 U.S. Population Census (Ruggles et al, 2010a), based on their anticipated age in 1880 (± 3 years), the phonetic proximity of their surname and given name (a value for the SPEDIS function in SAS ≤ 15; see Gershteyn (2000) for a description of the function)18, exact matches on the individual’s own birthplaces and the birthplaces of both parents, and only one person in both censuses meeting the preceding criteria. In order to reduce the probability of “false positives,” the linked dataset was further limited to individuals who either had a full given name (i.e. not just an initial) reported in both censuses and for whom no individual matching the other non-given name criteria was present in either census without a full given name.19 This process produced 398,181 matches from 1850 to 1880, a match rate of 13.0%.20                                                              See Xie & Killewald (2013) for a discussion of this issue.  The SPEDIS function is not symmetric in comparing two names a and b, so the value used was actually the average of SPEDIS(a,b) & SPEDIS(b,a). For example, for the surnames a=“Ferrie” and b=“Ferry,” SPEDIS(a,b)=22 and SPEDIS(b,a)=30, the average is 26 and the observation is rejected; for given names a=“Joseph” and b=“Joesph,” SPEDIS(a,b)=8 and SPEDIS(b,a)=8, the average is 8 and the observation is accepted if the other criteria are met. 19 For example, even if “Joseph Ferrie” in 1850 and “Joseph Ferrie” in 1880 were matched on age, own birthplace, and parents’ birthplaces, but there was also a “J. Ferrie” present in 1880 who matched on these criteria, the “Joseph Ferrie” match was rejected, on the basis of the possibility that the 1880 “J. Ferrie” was actually the correct “Joseph Ferrie” (because either the correct 1850 match for the “Joseph Ferrie” observed in 1880 was not successfully enumerated in 1850, or because the correct 1850 match for the “Joseph Ferrie” observed in 1880 was actually reported in 1850 as “J. Ferrie.” 20 There are several sources of loss in this linkage: the most likely are mis-enumeration (15%) or non-enumeration in either census (15%), age or name mis-reporting/changing (15% each), death between 1850 and 1880 (22.1% based on survival of this cohort inferred from the 1850 and 1880 population totals for males age 17 18

– 13 –

3. The sons of individuals matched 1850-1880 were then matched to the 1910 IPUMS 1% Sample (Ruggles et al., 2010b) using the same criteria as in Step (2). This yielded 1,886 unique matches. A random sample of 10,000 of the sons of 1850-1880 matches not found at this stage were then sought directly in the on-line 1910 U.S. Census index, of whom 5,019 were located in the 1910 census. We have transcribed a random subset consisting of 255 of these 5,019 matches to add to the first batch of 1,886 matches. The currently-transcribed data used here, then consists of 2,141 observations consisting of an occupation in 1850 for Generation 1 (grandfathers), an occupation in 1880 for Generation 2 (fathers), and a 1910 occupation for Generation 3 (sons). With the thirty-year gap between censuses, all observations occur at roughly the same point in each adult’s life cycle. Occupations have been grouped into the same four broad comparisons as used for Britain in the first, unranked classification scheme. There is no comprehensive measure of income by occupation for the total U.S. until the 1940 census, so occupations have instead been scaled by the average wealth (real estate & personal wealth) reported by white males age 24-55 in 1860 and 1870 for 106 distinct occupational titles. This will be referred to below as “occupational wealth.” Table 4 shows the characteristics of the linked data and of the general population meeting the linkage criteria. TABLE 4 REPRESENTATIVENESS OF LINKED SAMPLE, U.S.. 1850 1880 1910 Linked Pop Linked Pop Linked Age (years) Family Size Literate (%) ln(Occ Wealth) Region (%) Northeast Midwest South Far West N

Pop

37.0 7.5 93.4 7.7

35.7 6.2 92.9 7.6

37.2 6.5 99.3 7.6

35.5 5.5 99.2 7.5

36.4 5.7 97.5 7.4

35.7 7.0 96.8 7.3

50.8 13.3 35.7 0.3

42.2 26.6 29.6 1.6

42.1 24.1 32.3 1.4

30.9 37.6 27.4 4.1

42.2 23.3 33.5 1.0

28.3 40.5 28.2 3.0

33,308 1,473,882

35,392 3,060,761

43,163 6,714,240

Note: Males age 28-45 in each census linked across three generations with at least one co-resident male child. Family size reflects size only at the time of the census. See text for definition of occupational wealth. Regions defined by the Census Office.

                                                             3-21 in 1850 and age 33-51 in 1880). If we assume these factors are independent, the predicted match rate is 40.7%. Of these, half had at least one person sharing the same surname, given name, year of birth, birthplace, and parents’ birthplaces. The additional deletion of individuals with a potential match on criteria other than given name but an initial rather than a full given name reported in either census lowers this rate to 15%, just over the observed rate.

– 14 –

Estimating Multigenerational Effects Income and Wealth To test for the presence of a multigenerational effect on income and wealth, we follow the econometric framework of Lindahl et al. (2015). We estimate equations (1) and (2), where t represents G3, t-1 is G2, and t-2 is G1. Outcome variables are occupational earnings for Britain and occupational wealth for the U.S. As is common in the literature that measures IGE, the model specification is kept purposefully sparse in order that the β coefficients capture all the channels of intergenerational transmission. We compare (β1)2 to β2. In addition, we estimate ln

ln

,

ln

,

(5)

in order to directly assess the impact of grandfather’s earnings on grandson’s earnings, controlling for father’s earnings. It is important in what follows to keep in mind that our earnings and wealth results rely on imputations by occupation rather than direct observation. As our primary goal in the present study is simply to explore whether at least one generation prior to the paternal influenced individual economic outcomes historically in Britain and the U.S., rather than to place the magnitude of any of our results into any comparative framework from other contexts, these measures suffice. However, we cannot directly compare the magnitude of our estimates of intergenerational elasticity with the results from observed earnings that dominate the literature. A priori, it is not clear whether we should expect these results derived from imputed outcomes to be more or less likely to show multigenerational effects than would be the case if we observed real earnings. On one hand, there is reason to think that focusing on occupation rather than earnings would underestimate the impact of grandfathers. Suppose there is in reality a positive correlation between grandfather’s earnings and grandson’s, and that the correlation has both a within and a between-occupation component. So, for example, high-earning grandfathers transmit to grandsons both a tendency to hold a lucrative (on average) occupation and a tendency to earn more than average conditional on occupation. In this case, our data would understate the true degree of multigenerational correlation, as it would reflect only the between-occupation component and would reveal nothing about correlation within occupations. On the other hand, there could be conditions under which occupations show stronger multigenerational correlation than earnings. One way that grandfathers might matter is by revealing additional information about the father and/or the father’s household that is pertinent to the son’s outcome. In this way, grandfathers do not directly influ-

– 15 –

ence grandsons; rather, they add additional relevant information that essentially compensates for a sort of measurement error in equation 1. In a model that relies on occupation rather than earnings, there is more scope for such missing paternal information in that we cannot differentiate between high and low earning fathers within occupation. While theory does not offer a clear prediction on the relationship of an effect estimated on occupations versus earnings, a recent empirical study is relevant to the question. Feigenbaum (2017) exploits a relatively rich longitudinal data source from Iowa from 1915-1940 to show that measures of intergenerational mobility derived from earnings, occupation, and education were quite similar. While his results are confined to twentieth century Iowa, they do offer some evidence to believe that results derived from occupations and earnings might be similar in other contexts. Table 5 presents these results for Britain (Columns 1-3) and the U.S. (Columns 46). For both Britain and the U.S., the effect of both fathers and grandfathers on sons is economically and statistically significant when each regression is estimated separately (Columns 1, 2, 4, and 5). The effect of grandfathers’ earnings on grandsons’ remains significant when controlling for fathers’ earnings (Columns 3 and 6). The key result is the comparison of (β1)2 to β2. In both Britain and the U.S., it is clear that β2 > (β1)2, by a factor of two in both countries. Using the delta method, we can formally test the null nonlinear hypothesis β2 > (β1)2 after obtaining bounds on the standard errors of β2 and β12. Doing so yields 95% confidence intervals for the difference β2 – (β1)2 of [0.077, 0.103] for the U.S. and TABLE 5 REGRESSIONS OF PRIOR GENERATIONS' OUTCOMES ON SON'S OUTCOMES: Occupational Income (Britain) or Occupational Wealth (U.S.)

1 (Father)

(1) 0.269* (0.013)

Britain (2)

(3) 0.236* (0.014)

(4) 0.234* (0.005)

0.158* 0.066* (0.012) (0.013)

2 (Grandfather)

U.S. (5)

(6) 0.216* (0.006)

0.145* (0.006)

0.091* (0.006)

Constant

3.375

3.916

3.248

5.651

6.326

5.094

Obs

3,975

3,975

3,975

43,701

43,701

43,701

0.098

0.043

0.104

0.040

0.012

0.044

Adj. R

2

(1)2 2 2 - (1)

0.072

0.055

0.086 [0.059, 0.113]

0.090 [0.077, 0.103]

Note: Standard errors in parentheses. 95% confidence interval in brackets. * p-values < 0.01

– 16 –

[0.059, 0.113] for Britain, clearly excluding not just zero but also all negative values. In the regressions that include both fathers and grandfathers (Columns 3 and 6), the fathers’ outcomes are endogenous (determined solely by the grandfathers’ outcomes) if the underlying process is simply AR(1).21 Clearly, the evidence here strongly suggests that that is not the case. Grandfathers mattered to the occupational earnings or wealth realized by their grandsons over and above their influence on fathers’ outcomes. Although earnings and wealth information are not generally included in the censuses, the 1850 U.S. census does report the value of real estate holdings for household heads. With this information, we can begin to explore, for the U.S. at least, the mechanisms through which (dis)advantage is transmitted across multiple father-son transitions. If the regression in Column (6) of Table 5 is re-estimated, with the use of grandfather's actual 1850 real estate wealth (average value in the sample of $1799.33) in place of the grandfather’s wealth imputed on the basis of occupation, similar results are obtained: when a dummy for real estate > 0 is used, β2 = 0.119 (p = 0.045); when ln(real estate + $1) is used, β2 = 0.0189 (p = 0.018). The results in Table 5 invite comparison between Britain and the U.S., as we have done elsewhere (Long and Ferrie, 2013). However, the fundamental differences in the nature of the outcome variables precludes direct comparison. Partly for this reason, it is useful to test for multigenerational effects on occupational categories in addition to imputed earnings and wealth. With a harmonized occupational classification system, direct comparisons across countries are possible.

Occupation and Class In addition to yielding results appropriate for cross-country comparison, assessing multigenerational mobility in the framework of occupational classes in addition to imputed earnings and wealth offers several advantages. First and most importantly, as we do not directly observe earnings but rather impute them by occupation, it is important to demonstrate that our finding is robust to that method of treating our outcome variable. Second, dealing directly with occupational categories may be considered a more “direct” treatment of the occupational information that we actually observe, in that it requires fewer assumptions to translate occupation into a meaningful economic outcome variable. And finally, it places our findings in the context of class or socioeconomic mobility typically considered by social science disciplines outside of economics, in particular the large sociology literature on intergenerational mobility. With this in mind, we analyse general mo-

                                                             21 As it becomes possible to construct four-generation data for Britain and the U.S., we can account for this possibility by using the outcomes for great-grandfathers as an instrument for the outcomes of the fathers, as suggested by Becker and Tomes (1986).

– 17 –

bility patterns across Britain and the U.S. with the four-category occupational categorization scheme, and more nuanced upward and downward mobility patterns in Britain with the ranked five-category scheme. Table 6 shows occupational class mobility (transition probabilities) across three generations for Britain using the four-category unranked classification scheme, and Table 7 shows the same for the U.S. For both countries, three transitions are shown: grandfathers to fathers (G1  G2), fathers to sons (G2  G3), and grandfathers to (grand)sons (G1  G3). The occupational categories are not formally ranked, although moves from the unskilled category into any of the other three may unambiguously be considered upward moves while moves into the unskilled category may be considered downward moves. A striking feature of the mobility tables are the high rates of total mobility and of upward mobility out of unskilled occupations for the U.S. As we have shown elsewhere, mobility in the U.S. in this time period was higher than in Britain and higher than in the U.S. in the second half of the twentieth century (Long and Ferrie, 2013). The exception is the particularly high rate of persistence in farming between the first and second generation in the U.S. One particularly striking type of transition in the U.S. is the move into farming from other occupational backgrounds, which is far more common than in Britain, where movement into farming was very rare. It is not the case, however, that farming accounts for all of the mobility out of skilled, semiskilled, and unskilled occupations. There is also a great deal of mobility into white-collar occupations from these groups, and U.S., though more common than had traditionally been believed for the nineteenth century (Long, 2013). In Britain the occupationally mobile were likely to move into a skilled or semiskilled occupation, in large part because those occupations dominated the labour market at the time.22 In the context of occupational mobility, the central question is whether the influence of G1 on G3 is simply via the channel of G1’s influence on G2, or whether G1’s influence on G3’s occupational status is greater than this. To assess this question, we must compare G1-G3 mobility to mobility between G1-G2 and G2-G3 and calculate aggregate differences in total mobility across tables. To this end, we calculate the Altham statistic for each comparison across mobility tables. These are reported in Table 8, along with a comparison of each table to a matrix J of ones, representing perfect mobility; i.e. complete  

                                                             Long and Ferrie (2013) contains much more analysis of mobility in both countries, including comparisons that account for the relative frequency of different occupational categories.

22

– 18 –

TABLE 6

BRITISH OCCUPATIONAL MOBILITY ACROSS THREE GENERATIONS Column percents shown Generation 2 (Fathers, 1881) White Collar Farmer Skilled & SemiSkilled Unskilled

Generation 1 (Grandfathers, 1851) WC F S/SS U 36.0% 3.7 50.0 10.3

N 708 Total mobility = 45.8% Upward mobility = 53.4% Downward mobility = 14.1%

13.4% 46.4 25.5 14.7 388

15.5% 2.1 67.2 15.2 2,208

7.6% 2.1 43.8 46.6 1,639

Generation 2 (Fathers, 1881) WC F S/SS U

White Collar Farmer Skilled & SemiSkilled Unskilled

42.5% 2.0 44.5 11.0

Generation 3 (Sons, 1911) White Collar Farmer Skilled & SemiSkilled Unskilled

16.2% 28.3 35.0 20.4 314

16.3% 1.7 64.2 17.9 2,576

9.1% 2.1 51.1 37.7 1,244

N 568 Total mobility = 58.3% Upward mobility = 69.9% Downward mobility = 18.4%

4,943

N 904 173 2,734 1,076 4,887

(from Unskilled ) (to Unskilled )

Generation 1 (Grandfathers, 1851) WC F S/SS U 29.2% 1.9 54.8 14.1

773 287 2,654 1,229

(from Unskilled ) (to Unskilled )

Generation 3 (Sons, 1911)

N 753 Total mobility = 48.2% Upward mobility = 62.3% Downward mobility = 16.7%

N

18.7% 14.8 44.5 21.9 310

19.4% 2.3 59.2 19.1 1,774

(from Unskilled ) (to Unskilled )

 

– 19 –

14.0% 2.2 53.6 30.1 1,350

N 757 127 2,224 894 4,002

TABLE 7

U.S. OCCUPATIONAL MOBILITY ACROSS THREE GENERATIONS Column percents shown Generation 2 (Fathers, 1880)

Generation 1 (Grandfathers, 1850) WC F S/SS U

White Collar Farmer Skilled & SemiSkilled Unskilled

40.6% 32.7 19.4 7.3

N Total mobility = Upward mobility = Downward mobility =

9.7% 69.3 12.1 8.9

17.3% 32.2 38.0 12.5

2,717 28,264 8,988 42.8% 75.3% (from Unskilled ) 9.6% (to Unskilled )

8.1% 5,704 40.6 24,880 26.6 8,361 24.7 4,756 3,732

Generation 3 (Sons, 1910)

Generation 2 (Fathers, 1880) WC F S/SS U

White Collar Farmer Skilled & SemiSkilled Unskilled

64.4% 8.8 20.3 6.5

N Total mobility = Upward mobility = Downward mobility =

20.3% 48.5 16.9 14.2

30.7% 11.8 44.3 13.2

5,704 24,880 8,361 52.6% 73.9% (from Unskilled ) 12.9% (to Unskilled )

4,756

Generation 1 (Grandfathers, 1850) WC F S/SS U

White Collar Farmer Skilled & SemiSkilled Unskilled

48.0% 19.1 24.2 8.7

N Total mobility = Upward mobility = Downward mobility =

32.7% 20.1 33.3 13.9

2,717 28,264 8,988 62.3% 79.0% (from Unskilled ) 13.7% (to Unskilled )

 

– 20 –

43,701

N

16.7% 12,091 24.1 14,701 33.1 10,650 26.1 6,259

Generation 3 (Sons, 1910)

25.1% 40.3 20.5 14.1

N

43,701

N

20.1% 12,091 26.5 14,701 32.4 10,650 21.0 6,259 3,732

43,701

independence of rows and columns. The Altham statistic is ½

d

,

∑

∑

∑

∑

log

(6)

It is an aggregation of the differences between each cross-product ratio in tables P and Q, both with r rows and s columns, where p and q denote the individual elements of each table. It measures how far the association between rows and columns in table P departs from the association between rows and columns in table Q. A simple likelihood-ratio χ2 statistic G2 (Agresti, 2002, p. 140) with (r - 1)(s - 1) degrees of freedom can then be used to test whether the matrix Θ with elements θij = log(pij/qij) is independent; if we can reject the null hypothesis that Θ is independent, we essentially accept the hypothesis that d(P,Q) ≠ 0 so the degree of association between rows and columns differs between table P and table Q. The metrics d(P,J) and d(Q,J) reveal which of tables P and Q are closer to independence; i.e. in which mobility is greater. So if, for example, d(P,Q) ≠ 0 and d(P,J) < d(Q,J), we can conclude that table P shows greater mobility than does table Q. 23 The Altham statistics confirm the expectation that father-son (G1-G2 and G2-G3) occupational persistence is greater than grandfather-grandson (G1-G3) persistence, for both the U.S. and Britain. This is indicated by the relative proximity of the G1-G3 tables to independence (J) compared to the G1-G2 and G2-G3 tables. In both countries, overall two-generation mobility appears to have been fairly stable from 1850/51 to 1910/11. In terms of relative mobility – controlling for the relative frequency of occupations, as the Altham statistic does – Britain shows a modest increase in mobility and the U.S. a very small decline. These two-generation mobility tables do not directly reveal the extent to which the transmission of occupation is a non-Markov, multigenerational process. The results in Table 8 indicate that grandfather’s occupation influenced grandson’s, in that the G1-G3 tables are significantly different from independence for both countries. However, we would expect this to be the case even if intergenerational mobility was an AR(1) process as long as the degree of association between two generations was strong enough that mean reversion would not occur in three generations. We explicitly test for a multigenerational effect via a three-step procedure: 1. Calculate the matrix M1 of transition probabilities between G2 and G3, and matrix M2 between G1 and G3.

                                                             Note that this distance measure is multidimensional rather than linear; therefore it need not be the case that d(P,Q) = |d(P,J) – d(Q,J)|. For more on the Altham statistic and its use to compare mobility across tables see Altham (1970), Altham and Ferrie (2005), and Long and Ferrie (2013). 23

– 21 –

TABLE 8 MOBILITY COMPARISONS ACROSS GENERATIONS

Altham Statistics d(P,Q) Assessing Each Table's Distance From Independence and Its Distance from Other Tables Table Britain P: Grandfathers (1851) vs. Fathers (1881)

d(P,J)

5.13* 20.79*

20.79* 7.79*

Q: Grandfathers (1851) vs. Sons (1911) U.S. P: Grandfathers (1850) vs. Fathers (1880)

13.59*

14.49* 3.96*

Q: Fathers (1880) vs. Sons (1910) P: Fathers (1880) vs. Sons (1910)

d(P,Q)

23.91*

Q: Fathers (1881) vs. Sons (1911) P: Fathers (1881) vs. Sons (1911)

d(Q,J)

16.13*

16.13* 7.53*

Q: Grandfathers (1850) vs. Sons (1910)

9.03*

Note: d(X,J) is the distance from the row-column association in Table X to that under independence (in Table J, with 1's in each cell); d(X,Y) is the distance between the row-column asociation in Table X and that in Table Y. p-values for χ2 test of H0: d(X,Y)=0 * < 0.01

2. Use M1 to calculate [OccupationsG1]M1M1, the projected hypothetical mobility table that would have occurred from G1 to G3 if M1 had simply operated twice: from G1 to G2 and again from G2 to G3. 3. Calculate the three Altham statistics – d(P,J), d(Q,J) and d(P,Q) – necessary to compare mobility from this hypothetical mobility table M1M1 to the observed

– 22 –

mobility table M2. If M1M1 exhibits more total mobility than M2, that is the occupational mobility equivalent of (β1)2 < β2. Table 9 shows this comparison for both Britain and the U.S. For both countries, observed G1-G3 mobility (table P) is farther from independence than is projected mobility (table Q) and the distance between P and Q is relatively large and statistically significant. Clearly, grandfather’s occupational status mattered to the status attained by their grandson’s beyond its direct impact on father’s status. This analytical framework is useful for assessing a general grandfather effect on total mobility, but it does not reveal much about the nature of mobility, in particular with respect to upward or downward social mobility across multiple generations. For Britain, we can use the five-category Armstrong classification system to measure the effect of grandfather’s socioeconomic status on the nature of social mobility experienced by (grand)sons. Because this is an ordered classification scheme, any intergenerational change in occupational group can be considered as an upward or downward transition in socioeconomic status. Table 10 shows G2-G3 mobility, grouped by G1-G2 mobility. It compares father-son mobility according to whether the father’s status relative to the grandfather was the same (A), was higher (B), or was lower (C). Comparing the three mobility tables, it is clear that the father’s origin, in addition to his own occupational class attainment, influences the status of the son. Sons of fathers who were upwardly mobile were themselves much more likely to be downwardly mobile (43.9%) than were sons with similar origins whose fathers were in the same class as their grandfather (26.2%). On the other hand, sons of downwardly mobile fathers were more likely to be upwardly mobile (53.4%) than were sons of immobile fathers (25.2%). Some of these differences are purely mechanical. For example, sons of upwardly mobile fathers by definition could not have class V origin. So all sons in table B are at risk for a downward move, but only sons with class I-IV fathers are at risk in table A. A similar asymmetry holds for upward mobility between tables A and C. This component of the difference is small, however, as the adjusted mobility rates for table A show. It is worth noting that these differences do not represent “regression to the mean,” which refers to the fact that since intergenerational transmission is not perfect, the influence of fathers wage, class, etc. on descendants decreases over generations, as subsequent generations eventually regress to the mean. To the contrary, our finding indicates that prior generations exert influence beyond their own sons, a fact which implies a slower rate of mean regression than would otherwise be observed. In this sense, this result is comparable to the positive coefficient on β2 in columns (3) and (6) of Table 5.

– 23 –

TABLE 9 TESTING FOR GRANDFATHER EFFECT WITH MOBILITY TABLES Observed vs Projected Mobility, Column Percents Great Britain, 1851-1911 Q: Projected G1-G3 Mobility

P: Observed G1-G3 Mobility G3 (Sons, 1911)

G1 (Grandfathers, 1851) WC F S/SS U

[OccupationsG1] M1M1 G3 G1 (Grandfathers, 1851) (Sons, 1911) WC F S/SS U

White Collar Farmer Skilled/SS Unskilled

29.2% 1.9 54.8 14.1

18.7% 19.4% 14.0% 14.8 2.3 2.2 44.5 59.2 53.6 21.9 19.1 30.1

White Collar Farmer Skilled/SS Unskilled

26.6% 19.1% 19.3% 16.0% 2.4 9.4 2.2 2.4 53.8 50.1 58.1 56.9 17.2 21.5 20.3 24.8

N

568

310

N

568

1,774

1,350

310

1,774

1,350

d(Q,J) = 9.2** d(P,Q) = 4.7*

d(P,J) = 13.6**

United States, 1850-1910 Q: Projected G1-G3 Mobility

P: Observed G1-G3 Mobility G3 (Sons, 1910)

G1 (Grandfathers, 1850) WC F S/SS U

[OccupationsG1] M1M1 G1 (Grandfathers, 1850) G3 (Sons, 1910) WC F S/SS U

White Collar Farmer Skilled/SS Unskilled

48.0% 19.1 24.2 8.7

White Collar Farmer Skilled/SS Unskilled

50.6% 13.9 25.7 9.8

N

128

N

2,717

25.1% 40.3 20.5 14.1 28,264

32.7% 20.1 33.3 13.9 8,988

20.1% 26.5 32.4 21.0 3,732

30.5% 30.7 24.6 14.2 1,457

38.0% 16.8 32.2 13.0

30.1% 23.3 30.8 15.7

366

190

d(Q,J) = 6.5** d(P,Q) = 3.6**

d(P,J) = 9.0**

Note: M1 is the 44 matrix of transition probabilities between fathers (G2) and sons (G3). Projected G1-G3 mobility is the mobility that G3 would have epxerienced releative to G1 had the M1 transition probabilities been realized over both G2 and G3. d(X,J) is the distance from the row-column association in table X to that under independence (in Table J with 1's in each cell); d(X,Y) is the distance between the row-column association in table X and that in table Y. 2

p-values for χ text of H0: d(X,Y)=0 ** < 0.01 * < 0.05

Conclusions Using new nationally-representative data covering three generations in Britain and the U.S. from 1850-1911, we have found strong evidence that occupational mobility was not a simple AR(1) process across only two generations. Within the framework of a linear regression, grandfather’s earnings significantly influenced grandson’s earnings beyond the

– 24 –

TABLE 10 FATHER/SON OCCUPATIONAL TRANSITIONS BY FATHER'S MOBILITY Great Britain, Ranked Occupational Scale A. G2 in Same Class as G1 (Father Not Mobile): G2s Class, 1881 G3 Class, 1911 I II III IV V I: Professional 34.2% 2.6% 1.1% 0.7% 0.0% II: Intermediate 44.7 32.9 12.8 7.8 6.8 III: Skilled 18.4 43.9 64.4 42.1 55.4 IV: Semi-Skilled 0.0 14.9 7.6 29.3 9.5 V: Unskilled 2.6 5.7 14.2 20.2 28.4 N Total mobility: Upward mobility: Downward mobility:

38 228 1,038 605 74 51.5% 25.2% (26.2% omitting obs where G2=I) 26.2% (27.2% omitting obs where G2=V)

B. G2 in Higher Class than G1 (Father Upwardly Mobile): G2s Class, 1881 G3 Class, 1911 I II III IV V I: Professional 5.1% 3.2% 0.9% 0.0% II: Intermediate 39.2 23.5 10.2 3.7 III: Skilled 40.5 58.1 65.9 46.3 IV: Semi-Skilled 7.6 5.6 6.7 29.6 V: Unskilled 7.6 9.7 16.3 20.4 N Total mobility: Upward mobility: Downward mobility:

79 54.6% 10.7% 43.9%

341

569

54

Total 1.7% 14.0 54.0 15.0 15.3 1,983

Total 1.9% 16.4 60.4 7.6 13.7 1,043

C. G2 in Lower Class than G1 (Father Downwardly Mobile): G2s Class, 1881 G3 Class, 1911 I II III IV V Total I: Professional 0.0% 3.5% 0.4% 1.0% 1.6% II: Intermediate 30.8 10.5 11.1 8.2 10.2 III: Skilled 61.5 72.7 56.3 52.6 60.8 IV: Semi-Skilled 7.7 5.6 18.4 12.1 12.0 V: Unskilled 0.0 7.7 13.8 26.0 17.0 N Total mobility: 62.3% Upward mobility: 53.4% Downward mobility: 8.9%

13

  – 25 –

286

261

388

948

direct influence of father’s earnings. In addition, the effect of grandfather’s earnings on grandson’s earnings exceeds the effect implied by simply squaring the two-generation, father-son effect. Patterns of occupational class mobility reveal a similar effect, indicating that the influence of G1 on G3 is robust to the assumptions made in imputing earnings to occupation. In both countries, grandson’s occupational class was more closely related to grandfathers class than would be expected from the simple two generation father-son transition probabilities operating twice. A more nuanced, ordered occupational classification scheme revealed that in Britain, sons were more likely to be downwardly mobile if their fathers were themselves upwardly mobile, and vice versa. These empirical exercises reveal the existence of a “grandfather effect” but not its precise nature. In particular, it is not clear to what extent grandfathers might literally have impacted the occupations of their grandsons as opposed to simply providing additional relevant information to the econometrician that is empirically meaningful. On one hand, grandfathers might have impacted grandsons through their human capital formation, via financial assistance, through differential social networks, or through other direct channels. On the other, perhaps incorporating information regarding grandfather’s occupation or earnings into the grandson’s earnings regression simply provides more information regarding the grandson’s true socioeconomic initial state, typically measured by father’s occupation and earnings. In this sense it might be thought of simply as a proxy for relevant differences across fathers that are not captured by occupation and earnings. A carpenter who is the son of a physician, for example, might pass on to his son a significantly different endowment of genes, human capital, educational inputs, wealth, social network, et cetera than would a carpenter who is himself the son of a carpenter. The aim of this project is to define, identify, and quantify three-generation mobility effects in two historical economies. Future work will explore the nature of that effect, and in particular will attempt to understand to what extent the effect is direct – operating literally through grandfathers themselves – or indirect – operating via unobservable characteristics of the father and/or the father’s household. It should, for example, eventually be possible to ascertain which households have a surviving grandfather, and for which of them the grandfather is resident in the same town or parish. Because very few of our grandfathers were living in the same household as their adult sons in 1880/81, the grandfathers are mostly not in our current three-generation dataset. However, they could be recovered by linking them directly from the 1850/51 census into the 1880/81 census. Because of significant attrition at each census linkage, it is likely that we will also need to increase the size of our three-generation datasets to accomplish this additional linkage. However, whether the effect is direct or indirect (or both) does not change the essential conclusion: for both Britain and the U.S. in the second half of the nineteenth century, assessments of mobility based on two-generation estimates significantly overstate – 26 –

the true amount of intergenerational mobility. The persistence of earnings and occupational class across generations was greater than has previously been thought. Whether this is true more generally, for other countries or for these countries more recently, is a question these data cannot answer; however, these findings lend credence to the idea that our current understanding of economic and social mobility in many countries, based as it is on two-generation studies, could well systematically overstate mobility and understate the impact of family background on educational and occupational attainment. If this is true, a full assessment of the impact of family background on the occupational prospects of an individual must take into account at least two, and perhaps more, previous generations. With our current data, we cannot rule out the possibility that even more generations might matter. More research on the multigenerational nature of social mobility is needed to answer the question.

References Agresti, Alan. 2002. Categorical Data Analysis. Hoboken, New Jersey, Wiley. Altham, P. M. 1970. "The Measurement of Association of Rows and Columns for an R X S Contingency Table." Journal of the Royal Statistical Society 32: 63-73. Altham, P.M. and Joseph P. Ferrie. 2005. "Contingency Table Comparisons". Unpublished. Armstrong, W.A. 1972. "The Use of Information About Occupation". Nineteenth Century Society. E.A. Wrigley. Cambridge, Cambridge University Press: 191-310. Bailey, Martha, Morgan Henderson and Catherine Massey. 2017. "How Do Automated Linking Methods Perform? Evidence from the Life-M Project". Unpublished. Baines, Dudley. 1985. Migration in a Mature Economy: Emigration and Internal Migration in England and Wales, 1861-1900. Cambridge, Cambridge University Press. Becker, Gary S. and Nigel Tomes. 1979. "An Equilibrium Theory of the Distribution of Income and Intergenerational Mobility." Journal of Political Economy 87 (6): 1153-1189. Becker, Gary S. and Nigel Tomes. 1986. "Human Capital and the Rise and Fall of Families." Journal of Labor Economics 4 (3): S1-39. Björklund, Anders and Kjell G. Salvanes. 2011. "Chapter 3 - Education and Family Background: Mechanisms and Policies". Handbook of the Economics of Education. Stephen Machin Eric A. Hanushek and Woessmann Ludger, Elsevier. Volume 3: 201247. Black, Sandra E. and Paul J. Devereux. 2011. "Recent Developments in Intergenerational Mobility". Handbook of Labor Economics. Ashenfelter Orley and Card David. New York, Elsevier. Volume 4, Part B: 1487-1541. Braun, Sebastian Till and Jan Stuhler. 2016. "The Transmission of Inequality across Multiple Generations: Testing Recent Theories with Evidence from Germany." The Economic Journal. Campbell, Cameron and James Z. Lee. 2011. "Kinship and the Long-Term Persistence of Inequality in Liaoning, China, 1749-2005." Chinese Sociological Review 44 (1): 71-103. Clark, Gregory. 2014. The Son Also Rises : Surnames and the History of Social Mobility. Princeton, New Jersey, Princeton University Press.

– 27 –

Clark, Gregory and Neil Cummins. 2015. "Intergenerational Wealth Mobility in England, 1858–2012: Surnames and Social Mobility." The Economic Journal 125 (582): 61-85. Clark, Gregory and Neil Cummins. 2015. "Surnames and Social Mobility." Human Nature 25 (4): 517-537. Dribe, Martin and Jonas Helgertz. 2016. "The Lasting Impact of Grandfathers: Class, Occupational Status, and Earnings over Three Generations in Sweden 1815–2011." The Journal of Economic History 76 (4): 969-1000. Feigenbaum, James J. 2017. "Multiple Measures of Historical Intergenerational Mobility: Iowa 1915 to 1940." Economic Journal (forthcoming) Feinstein, Charles. 1988. "The Rise and Fall of the Williamson Curve." Journal of Economic History 48 (3): 699-729. Ferrie, Joseph P. 1999. Yankeys Now: Immigrants in the Antebellum United States, 18401860. New York, Oxford University Press. Gershteyn, Yefim. 2000. "Use of Spedis Function in Finding Specific Values". Proceedings of the Twenty-Fifth Annual SAS Users Group International Conference. Indianapolis. Jackson, R. V. 1987. "The Structure of Pay in Nineteenth-Century Britain." The Economic History Review 40 (4): 561-570. Levi, Leone. 1885. Wages and Earnings of the Working Classes. Report to Sir Arthur Bass, M.P. London,, J. Murray. Lindahl, Mikael, Mårten Palme, Sofia Sandgren-Massih and Anna Sjögren. 2014. "A Test of the Becker-Tomes Model of Human Capital Transmission Using Microdata on Four Generations." Journal of Human Capital 8 (1): 80-96. Lindahl, Mikael, Mårten Palme, Sofia Sandgren-Massih and Anna Sjögren. 2015. "Long-Term Intergenerational Persistence of Human Capital: An Empirical Analysis of Four Generations." Journal of Human Resources 50 (1): 1-33. Long, Jason. 2005. "Rural-Urban Migration and Socioeconomic Mobility in Victorian Britain." Journal of Economic History 65 (1): 1-35. Long, Jason. 2013. "The Surprising Social Mobility of Victorian England." European Review of Economic History 17 (1): 1-23. Long, Jason and Joseph P. Ferrie. 2013. "Intergenerational Occupational Mobility in Britain and the U.S. Since 1850". American Economic Review. 103. Mare, Robert D. 2011. "A Multigenerational View of Inequality." Demography 48 (1): 123. Mare, Robert D. and Xi Song. 2012. Social Mobility in Multiple Generations. Population Association of America Annual Meeting, San Francisco. Minnesota Population Center. 2015. "North Atlantic Population Project: Complete Count Microdata. Version 2.3 [Dataset]". Minneapolis, Minnesota Population Center. Mitchell, B. R. 1962. Abstract of British Historical Statistics. Cambridge, University Press. Peters, H. Elizabeth. 1992. "Patterns of Intergenerational Mobility in Income and Earnings." The Review of Economics and Statistics 74 (3): 456-466. Ruggles et al, Steven. 2010a. "Integrated Public Use Microdata Series: Version 5.0 [Machine-Readable Database]". Minneapolis: University of Minnesota. Ruggles, Steven, J. Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder and Matthew Sobek. 2010b. "Integrated Public Use Microdata Series: Version 5.0 [Machine-Readable Database]". Minneapolis: University of Minnesota.

– 28 –

Schürer, K. and M. Woollard. 2003. "National Sample from the 1881 Census of Great Britain [Computer File]". Colchester, Essex, History Data Service, UK Data Archive [distributor]. Schürer, K. and M. Woollard. 2008. "National Sample from the 1851 Census of Great Britain [Computer File]". Colchester, Essex, History Data Service, UK Data Archive [distributor]. Schürer, K. and M. Woollard. 2014. "National Sample from the 1911 Census of Great Britain [Computer File]". Colchester, Essex, History Data Service, UK Data Archive [distributor]. Solon, Gary. 1992. "Intergenerational Income Mobility in the United States." American Economic Review 82 (3): 393-408. Solon, Gary. 1999. "Intergenerational Mobility in the Labor Market". Handbook of Labor Economics. Orley Ashenfelter and David E. Card. New York, Elsevier. 3: 3 v. (xxi, 1277-3630 p.). Solon, Gary. 2004. "A Model of Intergenerational Mobility Variation over Time and Place". Generational Income Mobility in North America and Europe. Miles Corak. Cambridge, Cambridge University Press: 38-47. Solon, Gary. 2015. "What Do We Know So Far About Multigenerational Mobility?". NBER Working Paper. Steckel, Richard. 1988. "Census Matching and Migration: A Research Strategy." Historical Methods 21: 52-60. Warren, John Robert and Robert M. Hauser. 1997. "Social Stratification across Three Generations: New Evidence from the Wisconsin Longitudinal Study." American Sociological Review 62 (4): 561-572. Williamson, Jeffrey G. 1980. "Earnings Inequality in Nineteenth-Century Britain." Journal of Economic History 40 (3): 457-475. Williamson, Jeffrey G. 1982. "The Structure of Pay in Britain, 1710-1911." Research in Economic History 7: 1-54. Xie, Yu and Alexandra Killewald. 2013. "Intergenerational Occupational Mobility in Britain and the Us since 1850: Comment". American Economic Review. 103. Zimmerman, David J. 1992. "Regression toward Mediocrity in Economic Stature." American Economic Review 82 (3): 409-429.

– 29 –